Nine container images, the proprietary DocxIntel models, a Helm chart, SDKs, connectors, an offline documentation set and a security pack with an SBOM per image. Roughly 62 GB, signed, checksummed, and complete on arrival — because a bundle that needs to fetch something later is not an air-gapped bundle.
Bundle size varies by release as models are added or replaced. The exact manifest, with per-file sizes and SHA-256 digests, is published with every release and is available before you sign anything.
Each ships as its own OCI image with its own SBOM, so you can scan, sign off and scale them independently. Only four of them ever request a GPU.
TLS termination, OIDC and SAML token validation, role enforcement, rate limiting and request auditing. The only service anything outside the cluster ever talks to.
The job state machine. Splits documents into page tasks, enqueues them by priority class, tracks attempts and retries, and commits state to PostgreSQL before acknowledging anything.
Page normalisation, layout segmentation, reading-order resolution, table-structure reconstruction and text recognition across Arabic and Latin scripts, including handwriting.
Field, entity and clause extraction against schemas you define in the console, with every returned value carrying the page number and bounding box it came from.
Document type classification and automatic splitting of multi-document packets, so a 300-page submission arrives at your workflow as the twelve documents it actually contains.
Chunking, embedding and vector search over the corpora you choose to index, providing the grounded context that makes Ask answers citable back to a page.
Serves DocxIntel Reason and DocxIntel Rerank, the proprietary models behind Ask, drafting, comparison, mapping and summarisation, over an internal interface reachable only inside the cluster.
The browser UI: review queues with side-by-side page highlighting, schema and taxonomy design, user and role administration, audit search and the offline documentation set.
PDF rasterisation, format conversion, archive expansion, connector polling and result delivery. CPU-bound work kept off the GPU tier, and usually the first thing you scale.
Every model in the engine is built by BizfyLabs. They ship as files, land on a volume you own, are checksummed at load, and run on your GPUs — no third-party AI provider anywhere in the path.
No customer document, prompt or extracted value is ever used to train or fine-tune anything. Your data trains nothing, because your data never leaves your infrastructure.
One archive, one manifest, one checksum file. Everything below is inside it on the day it reaches you.
The security pack describes how the architecture supports your own certification and audit work. It does not claim certifications on our behalf — see the security page for exactly what is and is not asserted.
Most procurement delays are not disagreements. They are waiting — for an SBOM, for a data-flow diagram, for someone to answer question 47 of a vendor questionnaire. The security pack is in the bundle so that queue never forms.
BizfyLabs does not claim SOC 2 or ISO 27001 certification for DocxIntel. What the pack provides is the evidence your own certification and regulatory work needs — which, for an on-premise deployment inside your perimeter, is the thing that actually matters.
A cloud parsing subscription is a good product with a very short packing list. The comparison is not about quality — it is about what you are left holding.
What you receive
Model weights
Works with no internet
SBOM in the delivery
Documentation offline
Self-hosting available on
Cost of processing more pages
Comparison reflects publicly documented behaviour of metered document-parsing platforms as of 2026, including credit-based per-page pricing and enterprise-gated self-hosting. Verify against current vendor documentation before making a decision — we would rather you checked.
What happens to the artefacts after they arrive, and what you can rely on three years in.
The bundle arrives on encrypted media, through your transfer gateway, or as a download depending on your deployment model. You verify the cosign signatures and the SHA-256 manifest before anything is loaded — on your side of the boundary, with your own tooling.
Load the SBOMs into your scanner and the images into your registry. Because the SBOMs ship inside the bundle, an air-gapped security review needs no vulnerability-database lookup at delivery time beyond the feed you already mirror.
One Helm release brings up all nine services. The bundled validation job then checks image digests, model checksums, GPU availability and end-to-end processing on a known document set, so you know the install is sound before real traffic reaches it.
The bundled Grafana dashboards and Prometheus alert rules go straight into the monitoring stack you already run. Nothing is reported anywhere else — there is no vendor dashboard watching your cluster, because there is no connection for one.
Feature releases quarterly, LTS releases with an 18-month security window, and cumulative upgrades so a long-disconnected cluster moves forward in one step. Roll through a canary namespace first; the chart supports it out of the box.
If the relationship ends, the extracted data is already in your storage in open formats and the model weights are already on your volumes under their own licences. There is no export project, because there was never anything to export.
The bundle questions that come up in every technical review, answered specifically.
Nine signed OCI container images, the proprietary DocxIntel model set of roughly 46 GB, a Helm chart, a Docker Compose file for single-node sites, a CycloneDX SBOM per image, an offline documentation set and a security pack. That is the whole bundle — around 62 GB compressed. There is no hidden component that reaches out for anything else at runtime.
